Method, computing device and medium for enterprise risk assessment based on news text

Abstract

The present disclosure provides a method, computing device, and computer-readable storage medium for performing enterprise risk assessment based on news text. The method includes: determining whether the news text satisfies a predetermined condition; and, in response to determining that the news text satisfies the predetermined condition, determining that the enterprise has a public opinion risk, wherein determining that the news text satisfies the predetermined condition comprises at least two of the following: determining that the event risk value of the public opinion event corresponding to the news text is greater than a risk threshold; determining that the sentiment polarity of the news text is lower than a sentiment threshold; determining a topic category of the news text and determining that the determined topic category belongs to a specific risk topic set; and identifying a mentioned entity in the news text and determining that the mentioned entity hits the enterprise knowledge graph of the enterprise.

Description

Technical Field

[0001] The present disclosure generally relates to the field of machine learning, and more specifically, to a method, computing device, and computer-readable storage medium for performing enterprise risk assessment based on news text. Background Art

[0002] For many companies, especially those directly impacting people's livelihoods, public opinion significantly impacts their reputation. Negative public opinion, in particular, can severely or even devastatingly impact a company if not promptly addressed. In the current information age, with the advancement of internet technology, more and more news and information are appearing online instantly. With the flourishing of various online media, especially self-media, public opinion is spreading more rapidly and from more diverse sources. Therefore, relying solely on manual review by a company's marketing or public relations department is no longer sufficient to fully understand public opinion, especially to capture negative public opinion before it spreads widely.

[0003] Therefore, there is a need to automatically collect various public opinion messages through computer systems and conduct timely analysis to assess corporate risks and issue early warnings. Summary of the Invention

[0004] In response to at least one of the above problems, the present disclosure provides a solution for evaluating the risk impact of collected news texts (also referred to as public opinion in this article) on enterprises by analyzing multiple dimensions including emotions, topics, entities, relationships, events, etc.

[0005] According to one aspect of the present disclosure, a method for enterprise risk assessment based on news text is provided. The method includes: determining whether the news text satisfies a predetermined condition; and in response to determining that the news text satisfies the predetermined condition, determining that the enterprise has a public opinion risk, wherein determining that the news text satisfies the predetermined condition includes at least two of the following: determining that the event risk value of the public opinion event corresponding to the news text is greater than a risk threshold; determining that the sentiment polarity of the news text is lower than a sentiment threshold; determining the topic category of the news text and determining that the determined topic category belongs to a specific risk topic set; and identifying a mentioned entity in the news text and determining that the mentioned entity hits the enterprise knowledge graph of the enterprise.

[0006] In some embodiments, determining the event risk value of the public opinion event corresponding to the news text includes: establishing a public opinion event model to at least identify the event type, the event trigger words and event elements corresponding to each event type; using the public opinion event model to identify public opinion events from the news text, the public opinion events at least including the trigger words and event elements of the public opinion events; and using the event risk model to determine the event risk value of the public opinion event.

[0007] In some embodiments, using the public opinion event model to identify public opinion events from the news text includes: determining an event representation method of the news text, wherein the event representation method includes any one of a headline event representation, a central event representation, an event extraction representation, and an event semantic representation; based on the event representation method, using the public opinion event model to identify the public opinion event from the news text; and generating a short text public opinion event representation for the identified public opinion event.

[0008] In some embodiments, using an event risk model to determine the event risk value of the public opinion event includes: constructing a feature word set of the public opinion event based on the event type, event trigger words and event elements of the public opinion event; determining the average word vector of the feature word set; comparing the average word vector with multiple average word vectors in the parameters of the event risk model to determine the closest average word vector; and determining the risk value of the public opinion event based on the closest average word vector.

[0009] In some embodiments, the public opinion event model is further used to identify an event element role indicating a role of the event element and / or an event description sentence containing an event trigger word and an event element.

[0010] In some embodiments, the method further includes: constructing a historical event library associated with the enterprise, the historical event library containing multiple historical events with risk types of interest, each historical event being constructed based on the public opinion event model; labeling each historical event in the historical event library with a risk value, the risk value indicating the risk level of the historical event; training the event risk model using historical events with risk values ​​to generate a trained event risk model; and inputting the public opinion event into the trained event risk model to determine the event risk value of the public opinion event.

[0011] In some embodiments, determining that the sentiment polarity of the news text is lower than a sentiment threshold includes: determining the sentiment polarity of the news text using a sentiment classification model, the sentiment polarity indicating whether the sentiment training text has positive sentiment, negative sentiment, or neutral sentiment; and comparing the sentiment polarity of the news text with the sentiment threshold to determine that the sentiment polarity of the news text is lower than the sentiment threshold.

[0012] In some embodiments, the method further includes: constructing a sentiment training text set associated with the enterprise, the sentiment training text set containing multiple sentiment training texts, each sentiment training text including a news text associated with the enterprise; labeling sentiment polarity for each sentiment training text in the sentiment training text set; training a sentiment classification model using the sentiment training text and the labeled sentiment polarity to generate a trained sentiment classification model; and inputting the news text into the trained sentiment classification model to determine the sentiment polarity of the news text.

[0013] In some embodiments, determining the sentiment polarity of the news text using a sentiment classification model includes: filtering the text data contained in the news text; segmenting the filtered text data, and mapping the segmented text data into a text identifier sequence, wherein each text identifier sequence includes one or more discrete encoding vectors; inputting the generated text identifier sequence into the embedding layer of the sentiment classification model to convert the text identifier sequence into a continuous embedding vector; encoding the embedding vector output by the embedding layer in the encoding layer of the sentiment classification model to generate a contextual representation of the embedding vector; generating a sentence-level feature vector of the embedding vector using an attention function in the pooling layer of the sentiment classification model; generating a quality probability distribution of the sentence-level feature vector, the quality probability distribution indicating the probability of the news text being at different sentiment polarities; and selecting the sentiment polarity with the highest probability in the quality probability distribution as the sentiment polarity of the news text.

[0014] In some embodiments, labeling the sentiment polarity of each sentiment training text in the sentiment training text set includes: searching for a risk word in a risk word dictionary in the sentiment training text; if a risk word is found in the sentiment training text, determining the semantics associated with the risk word based on dependency syntactic rules; and labeling the sentiment training text based on the semantics associated with the risk word.

[0015] In some embodiments, determining the topic category of the news text includes determining the topic category of the news text using a topic classification model, the topic classification model includes a topic extraction layer and a topic clustering layer, and determining the topic category of the news text using the topic classification model includes: in the topic extraction layer, using a two-word topic model to determine the probability distribution of the title of the news text in a predetermined topic set and the probability distribution of each word in the title in each topic in the predetermined topic set; and in the topic clustering layer, based on the probability distribution of the title of the news text in the predetermined topic set, using a clustering algorithm to determine the topic category of the news text.

[0016] In some embodiments, the topic classification model further includes a topic description layer, and the method further includes: in the topic description layer, selecting central keywords from the news text to form a topic description of the news text.

[0017] In some embodiments, in the topic extraction layer, a two-word topic model is used to determine the probability distribution of the title of the news text in a predetermined topic set and the probability distribution of each word in the title in each topic in the predetermined topic set, including: segmenting the title of the news text to obtain multiple words of the title; organizing two words of the multiple words into a two-word pair and assigning a topic in the predetermined topic set to the two-word pair; determining the probability distribution of the topics of all two-word pairs composed of the multiple words; and determining the probability distribution of each word in the two-word pair based on the probability distribution of the topic of each two-word pair.

[0018] In some embodiments, at the topic clustering layer, based on the probability distribution of the title of the news text in a predetermined topic set, determining the topic category of the news text using a clustering algorithm includes: clustering the news titles into multiple clusters; identifying mentioned entities in the news text using an entity model; and merging the multiple clusters by similar clustering based on the mentioned entities in the news text to determine the topic category of the news text.

[0019] In some embodiments, at the topic description layer, selecting central keywords from the news text to constitute the topic description of the news text includes: extracting a summary of the news text; and selecting central keywords from the summary as the topic description based on the mentioned entities in the news text and the number of words in the news text.

[0020] In some embodiments, the method also includes: obtaining multiple historical topic descriptions in the predetermined topic set; performing text similarity calculation on the multiple historical topic descriptions to determine the text similarity between the multiple historical topic descriptions; and merging the topics in the predetermined topic set to update the predetermined topic set based on the text similarity between the multiple historical topic descriptions.

[0021] In some embodiments, an entity model is used to identify mentioned entities in the news text and determine whether the mentioned entities hit the enterprise knowledge graph of the enterprise, including: constructing an enterprise knowledge graph of the enterprise, the enterprise knowledge graph including multiple entities indicating business elements of the enterprise, entity attributes of each entity, and entity relationships between each entity; using the entity model to identify one or more mentioned entities, attributes of each mentioned entity, and entity relationships between each mentioned entity from the news text; determining the entity distance between the mentioned entity and multiple entities in the enterprise knowledge graph; and determining whether the mentioned entity hits the enterprise knowledge graph based on the entity distance.

[0022] In some embodiments, constructing the enterprise knowledge graph of the enterprise includes: constructing an industry knowledge graph based on the industry characteristics of the industry to which the enterprise belongs, wherein the industry knowledge graph includes a set of entities indicating various business elements of the industry, a set of entity attributes of each entity, and a set of entity relationships between various entities; and constructing the enterprise knowledge graph of the enterprise based on the business data of the enterprise and the industry knowledge graph.

[0023] In some embodiments, the method further comprises: constructing a public opinion entity for the news text, wherein the attributes of the public opinion entity include the topic category and sentiment polarity of the news text;

[0024] Add the public opinion entity and the mentioned entity to the enterprise knowledge graph.

[0025] In some embodiments, the entity model includes a serial entity recognition model and an entity relationship model, and the method further includes: labeling entities in the training text and entity relationships between entities; training the entity recognition model using the labeled training text to generate a trained entity recognition model; training the entity relationship model using the labeled training text to generate a trained entity relationship model; using the trained entity recognition model to identify one or more mentioned entities and attributes of each mentioned entity from the news text; and determining the entity relationships between the mentioned entities using the trained entity relationship model and one or more mentioned entities identified from the news text.

[0026] In some embodiments, the entity model includes a joint recognition model of entities and entity relationships, and the method further includes: labeling entities in the training text and entity relationships between entities; training the joint recognition model using the labeled training text to generate a trained joint recognition model; and using the trained joint recognition model to identify one or more mentioned entities and attributes of each mentioned entity and entity relationships between each mentioned entity from the news text.

[0027] In some embodiments, the joint recognition model is an entity and relationship joint recognition model based on parameter sharing.

[0028] In some embodiments, the entity and relationship joint recognition model based on parameter sharing includes an embedding layer, a shared encoding layer, an entity decoding layer, a relationship decoding layer, and a loss function layer, and the joint recognition model is trained using the annotated training text to generate a trained joint recognition model, including: inputting the annotated training text into the embedding layer to convert it into an embedding vector; extracting the semantic encoding of the training text from the embedding vector using a context model in the shared encoding layer, and the semantic encoding contains the grammatical and syntactic features of the training text; in the entity decoding layer, calculating the score distribution of the entity label category of the training text according to the semantic encoding; in the relationship decoding layer, calculating the score distribution of the relationship label category of the training text according to the semantic encoding; in the loss function layer, using the cross entropy function, based on the score distribution of the entity label category, the score distribution of the relationship label category and the true distribution of the training text, respectively, to determine the loss value; and iteratively optimizing the entity and relationship joint recognition model based on parameter sharing based on the loss value.

[0029] In some embodiments, the attribute of the entity mention includes the sentiment polarity of the entity mention.

[0030] In some embodiments, determining the entity distance between the mentioned entity and multiple entities in the enterprise knowledge graph includes: determining a word vector as the entity name distance based on the entity name of the mentioned entity and the entity name of an entity in the enterprise knowledge graph; determining multiple word vectors as entity attribute distances based on each attribute of the mentioned entity and multiple attributes of an entity in the enterprise knowledge graph; and determining the entity distance between the mentioned entity and an entity in the enterprise knowledge graph based on the entity name distance and the entity attribute distance.

[0031] In some embodiments, determining that the mentioned entity hits the enterprise knowledge graph based on the entity distance includes: determining whether the entity distance is less than a predetermined distance threshold; and if the entity distance is less than the predetermined distance threshold, determining that the mentioned entity hits the entity in the enterprise knowledge graph.

[0032] According to another aspect of the present disclosure, a computing device is provided. The computing device includes: at least one processor; and at least one memory, the at least one memory being coupled to the at least one processor and storing instructions for execution by the at least one processor, wherein when the instructions are executed by the at least one processor, the computing device performs the steps according to the above method.

[0033] According to yet another aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program code is stored. When the computer program code is executed, the method described above is executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present disclosure will be better understood and other objects, details, features and advantages of the present disclosure will become more apparent through the following description of specific embodiments of the present disclosure given with reference to the accompanying drawings.

[0035] Figure 1 A schematic diagram of a system for implementing a method for performing enterprise risk assessment based on news text according to an embodiment of the present disclosure is shown.

[0036] Figure 2 A flowchart of a method for performing enterprise risk assessment based on news text according to some embodiments of the present disclosure is shown.

[0037] Figure 3 A flowchart of a method for performing enterprise risk assessment based on news text according to other embodiments of the present disclosure is shown.

[0038] Figure 4A A flowchart of a method for determining an event risk value of a public opinion event according to some embodiments of the present disclosure is shown.

[0039] Figure 4B A more detailed flow chart of a block diagram for identifying public opinion events according to some embodiments of the present disclosure is shown.

[0040] Figure 4C A more detailed flowchart of a block diagram for determining an event risk value of a public opinion event using an event risk model according to some embodiments of the present disclosure is shown.

[0041] Figure 5A A flowchart of a method for determining the sentiment polarity of news text according to an embodiment of the present disclosure is shown.

[0042] Figure 5B A structural diagram of a sentiment classification model according to an embodiment of the present disclosure is shown.

[0043] Figure 6A A flowchart of a method for determining a topic category of a news text according to an embodiment of the present disclosure is shown.

[0044] Figure 6B A structural diagram of a topic classification model according to an embodiment of the present disclosure is shown.

[0045] Figure 6C A more detailed flow chart of a block for determining a probability distribution for each word in a title according to some embodiments of the present disclosure is shown.

[0046] Figure 6D A more detailed flowchart of a block for determining topic categories of news text using a clustering algorithm according to some embodiments of the present disclosure is shown.

[0047] Figure 6E A more detailed flow chart of a block for determining a topic expression of a news text according to some embodiments of the present disclosure is shown.

[0048] Figure 7A A flowchart of a method for identifying entity mentions in news text using an entity model according to an embodiment of the present disclosure is shown.

[0049] Figure 7B A further detailed flowchart of a block for identifying entity mentions and entity relationships from news text using an entity model according to some embodiments of the present disclosure is shown.

[0050] Figure 7C A further detailed flowchart of a block for identifying entity mentions and entity relationships from news text using an entity model according to some other embodiments of the present disclosure is shown.

[0051] Figure 7D A structural diagram of an entity and relationship joint recognition model based on parameter sharing according to an embodiment of the present disclosure is shown.

[0052] Figure 7E A detailed flowchart of a block diagram for training the joint recognition model using annotated training text according to some embodiments of the present disclosure is shown.

[0053] Figure 7F A more detailed flow chart of a block for determining entity distance according to an embodiment of the present disclosure is shown.

[0054] Figure 8A A schematic diagram of an enterprise knowledge graph according to an embodiment of the present disclosure is shown.

[0055] Figure 8B A schematic diagram of an industry knowledge graph according to an embodiment of the present disclosure is shown.

[0056] Figure 9 A block diagram of a computing device suitable for implementing embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0057] The preferred embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although preferred embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0058] In the following description, for the purpose of illustrating various invention embodiments, certain specific details are set forth to provide a thorough understanding of the various invention embodiments. However, those skilled in the relevant art will recognize that the embodiments may be practiced without one or more of these specific details. In other cases, well-known devices, structures, and techniques associated with the present application may not be shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0059] Unless the context requires otherwise, throughout the specification and claims, the word "comprise" and variations such as "include" and "have" should be construed in an open, inclusive sense, that is, should be interpreted to mean "including, but not limited to."

[0060] Reference throughout this specification to "one embodiment" or "some embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Thus, the appearances of "in one embodiment" or "some embodiments" in various places throughout this specification are not necessarily all referring to the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any manner in one or more embodiments.

[0061] In addition, the terms "first", "second", "third", "fourth", etc. used in the specification and claims are only used to distinguish objects for the sake of clarity of description, and do not limit the size or other order of the objects they describe.

[0062] Figure 1 FIG. 1 is a schematic diagram of a system 1 for implementing a method for performing enterprise risk assessment based on news text according to an embodiment of the present disclosure. Figure 1As shown in , the system 1 includes one or more data sources 10, a computing device 20 and a network 30. The data source 10 and the computing device 20 can exchange data through the network 30. Here, the data source 10 can be, for example, a device for providing various (filtered) news texts (also referred to herein as public opinion), such as a third-party data source that pushes various news texts (for example, various news, media articles, government notices, etc.) to the computing device 20, or it can be an original data source that is actively crawled by the computing device 20 through crawlers and other technologies. The computing device 20 can obtain these news texts in real time and process these news texts to assess corporate risks. The computing device 20 may include at least one processor 22 and at least one memory 24 coupled to the at least one processor 22, the memory 24 storing instructions 26 that can be executed by the at least one processor 22, and the instructions 26, when executed by the at least one processor 22, execute at least part of the method described below. The specific structure of the computing device 20 can be, for example, combined as follows Figure 9 As stated.

[0063] In the solution disclosed herein, an enterprise's internal and external knowledge is first organized and integrated to construct an enterprise knowledge graph. Then, using multiple deep learning models, a multi-dimensional analysis of news texts emerging online, including at least two of sentiment, themes, entities, relationships, and events, is performed to identify risks associated with the enterprise contained in the news texts. The system then searches and infers the news texts within the enterprise knowledge graph to assess whether the news texts pose risks to the enterprise and / or the level of risk. Finally, the system integrates risk and impact information to make early warning decisions.

[0064] Risk assessment of enterprises based on news texts can include two levels: one is to assess whether the news poses a public opinion risk to the enterprise, and the other is to assess the extent of the public opinion risk (i.e., risk value) caused by the news to the enterprise. These two aspects can be performed separately or in combination, that is, first assess whether the news poses a public opinion risk to the enterprise, and then further assess the extent of the risk posed by the news to the enterprise if it is determined that the news poses a public opinion risk to the enterprise. The following is combined with Figure 2 and Figure 3 Describe the specific solutions for these two aspects respectively.

[0065] Figure 2 FIG. 2 shows a flow chart of a method 200 for performing enterprise risk assessment based on news text according to some embodiments of the present disclosure. The method 200 may be performed by, for example Figure 1 The following takes the execution in the computing device 20 as an example, combined with Figures 1 to 9 Method 200 is described.

[0066] like Figure 2As shown in , method 200 includes box 210, where the computing device 20 can determine whether the acquired news text meets a predetermined condition, and in box 220, when it is determined that the news text meets the predetermined condition (i.e., when the judgment in box 210 is "yes"), the computing device 20 can determine that the enterprise has a public opinion risk. Here, the predetermined condition can be two or more of the event risk value, sentiment polarity, topic category, and enterprise knowledge graph hit situation. Specifically, as Figure 2 As shown in , determining that the news text meets the predetermined condition may include at least two of blocks 212 , 214 , 216 , and 218 .

[0067] In block 212 , the computing device 20 may determine that the event risk value of the public opinion event corresponding to the acquired news text is greater than a risk threshold.

[0068] Here, as mentioned above, the news text can be obtained by the computing device 20 from various data sources 10 in real time. The public opinion event refers to the event representation identified from the news text. The public opinion event can be linked to historical public opinion events to determine the event risk value of the public opinion event based on the event risk model. The method for determining the event risk value of the public opinion event is, for example, combined with the following: Figures 4A to 4C Those skilled in the art will appreciate that the present disclosure is not limited to Figures 4A to 4C The implementation details of method 400 are not disclosed. Instead, other models can be used to combine historical public opinion events with known risks to determine the event risk value of the public opinion event.

[0069] In box 214, the computing device 20 may determine that the sentiment polarity of the news text is lower than the sentiment threshold. In this article, the news text as a whole may have its sentiment polarity. For example, in the case of three categories, the sentiment polarity of the news text may include positive, neutral and negative. In the case of five categories, the sentiment polarity of the news text may include positive, positive, neutral, negative and negative. Alternatively, in some other embodiments, the sentiment polarity may be represented as a sentiment value within a continuous sentiment range, such as a value between [0,1]. Regression models, classification models, etc. may be used to determine the sentiment polarity of the news text. In this article, a classification model is used. Depending on the sentiment classification model used and the training sample, different sentiment polarity classification methods may be implemented. The method of determining the sentiment polarity of the news text using the sentiment classification model is, for example, combined with the following: Figures 5A to 5B Those skilled in the art will appreciate that the present disclosure is not limited to Figures 5A to 5B The implementation details of method 500 are not disclosed. Instead, other sentiment classification models may be used to determine the sentiment polarity of the news text.

[0070] Here, the sentiment threshold may be, for example, a specific sentiment polarity or a specific sentiment value. For example, when using a three-category or five-category sentiment polarity, the sentiment threshold may be "neutral." When using a continuous sentiment range (e.g., [0, 1]), the sentiment threshold may be 0.5. In these cases, a sentiment polarity below the sentiment threshold may indicate that the sentiment polarity of the news text is biased towards negativity.

[0071] At block 216 , computing device 20 may determine a topic category of the news text and determine that the determined topic category belongs to a specific risk topic set.

[0072] Here, topic categories refer to different topics presented in news texts. Determining topic categories helps to quickly filter out the current news text from a large amount of news texts to see if it is of the type of news text that the company is interested in (for example, food safety). The topic classification model can be used to determine the topic category of the news text, for example, by combining the following Figures 6A to 6E Those skilled in the art will appreciate that the present disclosure is not limited to Figures 6A to 6E The implementation details of method 600 are not disclosed. Instead, other topic classification models may be used to determine the topic type of the news text.

[0073] A specific risk topic set refers to a set of topic types that an enterprise is interested in. For example, for a food company, the topic types that it is interested in may include food safety, food quality, food carcinogenicity risks, and other topic types that may pose a public opinion risk to the enterprise. In this case, the enterprise can set the topic types that it is interested in to the specific risk topic set. In box 216, after determining the topic category of the news text, it can be determined whether the topic category belongs to the specific risk topic set. If it is determined that the news text belongs to the specific risk topic set, then the news text is more likely to pose a public opinion risk to the enterprise.

[0074] In block 218 , the computing device 20 may identify entity mentions in the news text and determine whether the entity mentions hit the enterprise knowledge graph of the enterprise.

[0075] Generally, an entity refers to a person's name, place name, organization name, etc. that appears in a news text. Furthermore, an entity may also include any other subject of interest, such as a product name, material name, brand name, etc. Each entity may include one or more entity attributes. In addition, various entity relationships may exist between entities. In this article, an entity model can be used to at least identify the entities in the news text, the attributes of the entity, and the entity relationships between entities. In order to distinguish the entities identified from the news text from the entities displayed in the enterprise knowledge graph, the former are also referred to as "mention entities" in this article.

[0076] The method of using entity model to identify entities in news text is as follows Figures 7A to 7F Those skilled in the art will appreciate that the present disclosure is not limited to Figures 7A to 7F The implementation details of the method 700 are not described herein. Instead, other entity models may be used to identify entities in the news text.

[0077] If at least two of blocks 212, 214, 216, and 218 are satisfied, computing device 20 may determine in block 220 that the news text will create a public opinion risk for the enterprise. Otherwise, in block 230, computing device 20 may determine that the news text will not create a public opinion risk for the enterprise.

[0078] Figure 3 FIG. 3 is a flowchart of a method 300 for performing enterprise risk assessment based on news text according to another embodiment of the present disclosure. The method 300 may be performed by, for example Figure 1 The following takes the execution in the computing device 20 as an example, combined with Figures 1 to 9 Method 300 is described.

[0079] In block 310, the computing device 20 may construct a public opinion entity for the news text. The attributes of the public opinion entity include the topic category and sentiment polarity of the news text. Here, the method for determining the topic category of the news text may be, for example, combining Figures 6A to 6E The method 600 described above, the method for determining the sentiment polarity of a news text may be, for example, the method 500 described in conjunction with FIG5 . It will be appreciated by those skilled in the art that the present disclosure is not limited to the method 500 described in FIG5 . Figures 5A to 5B Method 500 and Figures 6A to 6E The implementation details of method 600 are not disclosed. Instead, other sentiment classification models may be used to determine the sentiment polarity of the news text and / or other topic classification models may be used to determine the topic type of the news text.

[0080] In block 320 , the computing device 20 may determine a direct relationship risk value and an associated risk value of the news text to the enterprise based on the public opinion entity and the enterprise knowledge graph of the enterprise.

[0081] In some embodiments, determining the direct relationship risk value of the news text to the enterprise based on the public opinion entity and the enterprise knowledge graph of the enterprise may include: using an entity model to identify the mentioned entity in the news text; determining the hit entity of the mentioned entity in the enterprise knowledge graph of the enterprise; and determining the direct relationship risk value between the public opinion entity and the hit entity based on the enterprise knowledge graph.

[0082] Here, the method of using entity model to identify entities in news text is combined with the following Figures 7A to 7FThose skilled in the art will appreciate that the present disclosure is not limited to Figures 7A to 7F The implementation details of the method 700 are not described herein. Instead, other entity models may be used to identify entities (also referred to as mention entities in this document) in the news text.

[0083] There may be a corresponding relationship between the entities identified from the news text and the entities in the enterprise knowledge graph. The entities identified from the news text may be subsidiaries, abbreviations, nicknames, etc. of a certain enterprise, while the entities in the enterprise knowledge graph may be standard names set for the enterprise. In this case, it is possible to determine whether the entity identified from the news text hits the entity in the enterprise knowledge graph based on the entity attributes. For example, suppose the entity "KFC" representing the enterprise name is extracted from a news text, and the entity "Kentucky" representing the enterprise name exists in the enterprise knowledge graph, and its entity attributes include the nickname "KFC", then it can be determined that the entity "KFC" identified from the news text hits the entity "Kentucky" in the enterprise knowledge graph.

[0084] In some embodiments, the risk value R path It can be determined by determining the total number of paths between the public opinion entity and the hit entity.

[0085] R path =∑Q(A, E i )

[0086] Among them, A is the public opinion entity, Ei is the hit entity, and Q(A,Ei) represents the number of paths from the public opinion entity A to the hit entity Ei.

[0087] In some embodiments, the direct relationship risk value R can be determined based only on the total number of paths within a predetermined distance (e.g., 2 or 3) between the public opinion entity and the hit entity. path .

[0088] For example, suppose the enterprise knowledge graph of the enterprise is as follows Figure 8A As shown, further, the construction of the enterprise knowledge graph 810 can be as follows Figure 7A As described in block 710 of Figure 8AAs shown in , it is assumed that the enterprise knowledge graph 810 includes entities E1-E9, the public opinion entity is A, and the entities identified from the news text (i.e., mentioned entities) include entities A1, A2, and A3. The mentioned entities A1, A2, and A3 hit entities E2, E7, and E9 respectively in the enterprise knowledge graph, and only the case where the distance between the public opinion entity and the hit entity is less than the predetermined distance 3 is calculated. Then, the number of paths between public opinion entity A and hit entity E2 can be expressed as Q(A,E2)=1 (i.e., there is only one path A—>A2—>E2 from public opinion entity A to hit entity E2), the number of paths between public opinion entity A and hit entity E7 can be expressed as Q(A,E7)=1 (i.e., there is only one path A—>A1—>E7 from public opinion entity A to hit entity E7), and the number of paths between public opinion entity A and hit entity E9 can be expressed as Q(A,E9)=2 (i.e., there are two paths A—>A2—>E2—>E9 and A—>A3—>E9 from public opinion entity A to hit entity E9). Therefore, the direct relationship risk value R path =1+1+2=4.

[0089] In addition, in some embodiments, the direct relationship risk value R path The calculation of may also consider the directionality of the paths between entities in the enterprise knowledge graph 810. For example, some paths among entities E1-E9 are unidirectional, while others are bidirectional.

[0090] In some embodiments, determining the associated risk value of the news text to the enterprise based on the public opinion entity and the enterprise knowledge graph 810 of the enterprise may include: based on a label propagation algorithm, determining the attention entities associated with the public opinion entity from the enterprise knowledge graph 810, and marking a risk value for each attention entity; and determining the associated risk value of the public opinion entity based on the risk values ​​of these attention entities.

[0091] Throughout the risk assessment process, the enterprise knowledge graph is constantly updated to incorporate newly discovered risk factors. In some cases, although the text directly identified from the news text may not directly hit the entity in the enterprise knowledge graph, there may be entities in the enterprise knowledge graph that are related to the news text and have potential risks and require special attention. In this case, the label propagation algorithm can be used to mine entities that may be associated with public opinion entity A from the enterprise knowledge graph (also referred to as attention entities here), and label each attention entity with a risk value. For example, a risky attention entity is labeled with a risk value of 1, and a risky attention entity that does not have a risk is labeled with a risk value of 0.

[0092] In this way, the associated risk value R of the public opinion entity A can be determined based on the risk values ​​of these concerned entities. tag .

[0093] R tag =∑T(E i )

[0094] Wherein, Ei is the entity of interest, and T(Ei) represents the risk value of the entity of interest (for example, 0 or 1).

[0095] In block 330 , the computing device 20 may determine an indirect relationship risk value of the news text to the enterprise based on the enterprise knowledge graph and the historical news text collection.

[0096] In some embodiments, the computing device 20 can determine multiple historical public opinion entities similar to the public opinion entity of the news text from the historical news text collection, and determine the indirect relationship risk value between these historical public opinion entities and the hit entity in the enterprise knowledge graph based on the enterprise knowledge graph. For example, the indirect relationship risk value R sim It can be determined by the similarity between the public opinion entity and the hit entity.

[0097] R sim =ΣS(A,E i )

[0098] Where A is the public opinion entity, Ei is the hit entity, and S(A,Ei) represents the similarity between the public opinion entity A and the hit entity Ei. Here, the similarity can be represented by the vector distance between the embedding vectors of the public opinion entity and the hit entity.

[0099] In some embodiments, for example, a graph embedding method (GEM), also known as network representation learning (NRL), can be used to represent each entity in the enterprise knowledge graph using a low-dimensional, dense, real-valued embedding vector to obtain an embedding matrix for the entire enterprise knowledge graph. Based on this embedding matrix, a vector representation of each input entity can be found, and then similarity metrics such as cosine similarity, Pearson correlation coefficient, KL divergence, and Jaccard similarity coefficient can be used to obtain the similarity between the input entities.

[0100] At block 340 , the computing device 20 may determine a risk value of the news text to the enterprise based on the direct relationship risk value and the associated risk value determined at block 320 and the indirect relationship risk value determined at block 330 .

[0101] In some embodiments, without considering the different impacts of direct relationship risk value, indirect relationship risk value and associated risk value on the public opinion risk of the enterprise, the direct relationship risk value, indirect relationship risk value and associated risk value can be directly added to obtain the risk value of the news text to the enterprise.

[0102] In other embodiments, considering the different impacts of direct risk value, indirect risk value, and associated risk value on the public opinion risk of an enterprise, different weights are set for the direct risk value, indirect risk value, and associated risk value, and the weighted sum of these values ​​is used to obtain the risk value of the news text to the enterprise. For example, the risk value of the enterprise can be expressed as:

[0103] R=λ path *R path +λ sim *R sim +λ tag *R tag

[0104] Among them, λ path ,λ sim and λ tag They are the weight values ​​of direct relationship risk value, indirect relationship risk value and associated risk value, which can be set based on experience or determined through experiments.

[0105] Figure 4A A flowchart of a method 400 for determining an event risk value of a public opinion event according to some embodiments of the present disclosure is shown.

[0106] like Figure 4A As shown in , in box 410, the computing device 20 can establish a public opinion event model. The public opinion event model can be used to at least identify event types, event trigger words corresponding to each event type, and event elements. Here, the public opinion event model can be established for the enterprise or any enterprise of interest. A public opinion event refers to an event representation of a news text (or public opinion). A public opinion event includes at least a trigger word and an event element. Among them, the trigger word is a word used to trigger the public opinion event, which can clearly express the occurrence of the event, usually a noun or verb, and the event element is the information involved in the event, which is related to the event representation method.

[0107] Furthermore, in some embodiments, in addition to trigger words and event elements, a public opinion event may also include at least one of an event type, an event element role, and an event description sentence. The event type indicates the type of public opinion event, the event element role indicates the role each event element plays in the public opinion event, and the event description sentence is a sentence describing the public opinion event, typically including trigger words and event elements.

[0108] Here, the public opinion event model is built for the company. The public opinion events, as well as the event types, trigger words, event elements, event element roles, and event descriptions contained within them, have values ​​relevant to the company. For example, Table 1 below lists the public opinion event types and corresponding trigger words (list) that are of concern to food companies.

[0109] Table 1

[0110]

[0111] In block 420 , the computing device 20 may use the public opinion event model to identify public opinion events from the news text.

[0112] Figure 4B A more detailed flow chart of block 420 for identifying public sentiment events according to some embodiments of the present disclosure is shown.

[0113] like Figure 4B As shown, block 420 may further include block 422, where the computing device 20 may determine an event representation of the news text. The event representation may include any one of a headline event representation, a central event representation, an event extraction representation, and an event semantic representation.

[0114] In block 424, the computing device 20 may identify public opinion events from the news text using the public opinion event model established for the enterprise based on the event representation. Here, identifying public opinion events from the news text using the public opinion event model may be done by identifying the trigger words and event elements of the public opinion event from the news text. Those skilled in the art will appreciate that the present invention is not limited to this, and the public opinion event model may also be constructed based on artificial intelligence algorithms or machine learning algorithms to identify public opinion events. Furthermore, for headline event representation, the public opinion event may be identified from the title of the news text. For central event representation, core words may be extracted from the title and body of the news text, and the public opinion event may be summarized from these core words. For event extraction representation, trigger words may be extracted from the news text, corresponding arguments may be found based on the trigger words, and the public opinion event may be summarized. For event semantic representation, after the event is summarized by the event extraction representation, the event may be transformed into a vector representation, and event fusion may be performed to obtain the public opinion event.

[0115] At block 426, the computing device 20 may generate a short text representation of the public opinion event identified in block 424. Here, the short text representation of the public opinion event refers to a short text representation of the identified public opinion event. Those skilled in the art will appreciate that the short text representation here is relative to the long text representation and may vary depending on the application scenario, for example, ranging from dozens to hundreds of characters.

[0116] For example, for the title event representation, some text preprocessing can be performed on the news text, such as traditional-simplified conversion, word segmentation, and removal of stop words, to generate a short text representation of the public opinion event. For the central event representation, the news text can be processed using methods such as Figure 6BThe topic classification model 650 is shown to obtain a short text representation of the public opinion event.

[0117] continue Figure 4A In block 430 , the computing device 20 may determine an event risk value of the public opinion event using an event risk model.

[0118] Figure 4C A more detailed flowchart of block 430 for determining an event risk value of a public opinion event using an event risk model according to some embodiments of the present disclosure is shown.

[0119] like Figure 4C As shown, block 430 may further include block 432, where the computing device 20 may construct a historical event library associated with the enterprise. The historical event library may contain multiple historical events with a risk type of interest, each of which is constructed based on the public opinion event model.

[0120] In block 434 , the computing device 20 may label each historical event in the historical event library with a risk value, where the risk value indicates the risk level of the historical event.

[0121] At block 436 , the computing device 20 may train the event risk model using historical events with risk values ​​to generate a trained event risk model.

[0122] At block 438 , the computing device 20 may input the public opinion event identified at block 420 into a trained event risk model to determine an event risk value for the public opinion event.

[0123] In some embodiments, the event risk model can be constructed based on the KNN (K-Nearest Neighbor) model. The KNN model is a traditional model whose classification principle is to classify any new sample into the category with the most categories among the K samples closest to the sample. Here, the event risk is marked as different levels of classes by using the frequency of occurrence of feature word events in the historical event library, and the corresponding event risk value is determined by determining which category the public opinion event corresponding to the news text belongs to. Of course, those skilled in the art will understand that the event risk model in the present invention can also be constructed based on other appropriate machine learning models.

[0124] Specifically, training the event risk model in box 436 may include: constructing a feature word set of the historical event based on the event type, event trigger word and event element of each historical event; determining the occurrence frequency of multiple feature word sets constructed for multiple historical events in the historical event library; selecting several feature word sets with the highest occurrence frequency from the multiple feature word sets, and determining the average word vector of each selected feature word set; finally, determining the parameters of the event risk model based on the average word vector to obtain the trained event risk model.

[0125] In this case, determining the event risk value of the identified public opinion event in block 438 may include: constructing a feature word set of the public opinion event based on the event type, event trigger words, and event elements of the public opinion event; determining an average word vector of the feature word set; comparing the average word vector with multiple average word vectors in the parameters of the event risk model to determine the closest average word vector; and determining the event risk value of the public opinion event based on the closest average word vector. In other words, the risk value of the historical event corresponding to the feature word set corresponding to the closest average word vector can be considered the risk value of the public opinion event.

[0126] use Figures 4A to 4C The method for determining the event risk value of a public opinion event shown can construct the acquired news text into an overall public opinion event, and the event risk model determines the correlation between the current news text and historical public opinion events, thereby determining the overall event risk value accordingly.

[0127] Figure 5A A flowchart of a method 500 for determining the sentiment polarity of a news text using a sentiment classification model according to an embodiment of the present disclosure is shown.

[0128] like Figure 5A As shown in , the method 500 may include block 510, wherein the computing device 20 may construct a sentiment training text set associated with the enterprise. The sentiment training text set includes a plurality of sentiment training texts, each of which includes a news text associated with the enterprise.

[0129] In block 520 , the computing device 20 may label each emotion training text in the emotion training text set with an emotion polarity. Here, taking the three-category emotion polarity as an example, the labeled emotion polarity may indicate whether the corresponding emotion training text has positive emotion, negative emotion, or neutral emotion.

[0130] In block 530 , the computing device 20 may train the sentiment classification model using the sentiment training text and the labeled sentiment polarity to generate a trained sentiment classification model.

[0131] In block 540 , the computing device 20 may input the current news text into the trained sentiment classification model to determine the sentiment polarity of the news text.

[0132] In some embodiments, the sentiment classification model can be built based on a BERT (Bidirectional Encoder Representation from Transformers) model. Figure 5B FIG. 5 shows a schematic diagram of the structure of the sentiment classification model 550 according to an embodiment of the present disclosure. Figure 5B As shown, the sentiment classification model 550 may include an embedding layer 551, an encoding layer 552, a pooling layer 553, and an output layer 554. Furthermore, a data preprocessing process may be included before the embedding layer 551 to generate a text identifier sequence input to the embedding layer, as described below. Of course, those skilled in the art will appreciate that the sentiment classification model in the present invention may also be constructed based on other appropriate machine learning models.

[0133] More specifically, training the sentiment classification model 550 in block 530 may further include filtering the text data contained in the sentiment training text, such as performing traditional / simplified Chinese conversion, special word processing, and / or filtering of predetermined tag formats within the site. The filtered text data is then segmented, and the segmented text data is mapped into a text identifier sequence, wherein each text identifier sequence includes one or more discrete encoding vectors. Here, the text identifier sequence can be generated, for example, using a dedicated tokenizer of the BERT model, which will not be described in detail here. The text identifier sequence may include a token encoding, a position encoding, a sentence position encoding, and an attention mask. The token encoding is used to indicate the encoding vector converted for each token in the text identifier sequence, the position encoding indicates the position of each token, and the sentence position encoding indicates which sentence each token belongs to (e.g., "0" indicates that the current token belongs to the first sentence, "1" indicates that the current identifier belongs to the second sentence, etc.). The attention mask indicates whether the current token is filled (e.g., "1" indicates that the current token is non-filled, "0" indicates that the current token is filled). By converting the sentiment training text into a text label sequence, the data preprocessing process before training is realized.

[0134] The generated text identifier sequence is input to the embedding layer 551 of the sentiment classification model 550 to convert the text identifier sequence into a continuous embedding vector. Then, in the encoding layer 552 of the sentiment classification model 550, the embedding vector output by the embedding layer is encoded to generate a contextual representation of the embedding vector. The encoding layer 552 can extract the output features of the embedding vector output by the embedding layer 551 and can be composed of a plurality of hidden layers with the same structure but different parameters connected in series. Each hidden layer includes, for example, a linear transformation, an activation function, a multi-head self-attention and a skip connection.

[0135] Then, in the pooling layer 553 of the sentiment classification model 550, the attention function can be used to generate a sentence-level feature vector of the embedding vector; and the quality probability distribution of the sentence-level feature vector is generated in the output layer 554 as the output vector of the sentiment training text. For example, the Softmax model can be used to obtain the quality probability distribution. The model convergence parameters of the sentiment classification model can be determined based on the output vector of the sentiment training text, the marked sentiment polarity and the loss function. For example, the forward propagation algorithm or the backpropagation algorithm can be used to update the weight functions of each layer of the sentiment classification model to obtain stable model parameters.

[0136] In some embodiments, in box 520, the sentiment training text can be annotated based on the risk word dictionary. Specifically, the risk words in the risk word dictionary can be searched in the sentiment training text. Here, the risk word dictionary can be a pre-set collection of words (i.e., risk words) that represent the risks or negative impacts within the industry. If a risk word is found in the sentiment training text, the semantics associated with the risk word can be determined based on the dependency syntactic rules. Dependency syntactic rules refer to the dependency relationship between words. For example, if a word modifies another word, the modifier is considered to be a subordinate word, and the modified word is called a dominant word. The relationship between the two is called a dependency relationship. Usually, the modified word and the modifier are combined to truly determine the true semantics of the modified word in the article. Then, the sentiment training text can be annotated based on the semantics associated with the determined risk words. For example, it can be annotated as positive emotion, neutral emotion, or negative emotion.

[0137] When the BERT model is used to construct the sentiment classification model 550, determining the sentiment polarity of the news text in block 540 may include filtering the text data contained in the news text, such as performing traditional Chinese to simplified Chinese conversion, special word processing, and / or filtering for predefined tag formats within the website. Then, the filtered text data is segmented, and the segmented text data is mapped into a text identifier sequence, where each text identifier sequence includes one or more discrete encoding vectors.

[0138] The generated text token sequence is input into the embedding layer 551 of the sentiment classification model 550 to convert the text token sequence into a continuous embedding vector. The embedding vector output by the embedding layer 551 is then encoded in the encoding layer 552 of the sentiment classification model 550 to generate a contextual representation of the embedding vector.

[0139] Then, in the pooling layer 553 of the sentiment classification model 550, an attention function is used to generate a sentence-level feature vector for the embedded vector, and a softmax model is used in the output layer 554 to generate a quality probability distribution for the sentence-level feature vector. The resulting quality probability distribution indicates the probability of the news text being at different sentiment polarities. Therefore, the sentiment polarity with the highest probability in the quality probability distribution can be selected as the sentiment polarity of the news text.

[0140] use Figure 5A The method of determining the sentiment polarity of a news text using the sentiment classification model 550 shown can analyze the sentiment of the news text as a whole, and thus can evaluate the risk situation of the public opinion represented by the news text to the enterprise based on the overall sentiment of the news text.

[0141] Although the above combined Figure 5A The sentiment classification model according to the present disclosure is described using the BERT model as an example. However, those skilled in the art will appreciate that the present disclosure is not limited thereto, and deep learning models such as TextCNN and Attention-BILSTM can also be used to construct a sentiment classification model that implements full-text level sentiment classification.

[0142] In addition, in some embodiments, it is also possible to construct an entity-level sentiment classification (for example, in combination with Figures 6A to 6E After the entity recognition described in method 600), the entity-level sentiment classification model is combined with Figure 5B The full-text level sentiment classification model 550 may adopt the same model structure but use different training samples.

[0143] Figure 6A A flowchart of a method 600 for determining a topic category of a news text according to an embodiment of the present disclosure is shown. Figure 6B FIG. 6 shows a schematic diagram of the structure of a topic classification model 650 according to an embodiment of the present disclosure. Figure 6B As shown, the topic classification model 650 may include a topic extraction layer 651 and a topic clustering layer 652. In some embodiments, the topic classification model 650 may further include a topic description layer 653.

[0144] like Figure 6AAs shown in , in box 610, in the topic extraction layer 651, the computing device 20 can use a biterm topic model (BTM) to determine the probability distribution of the title of the news text in a predetermined topic set and the probability distribution of each word in the title in each topic in the predetermined topic set.

[0145] The BTM model can learn topics by modeling the generation of bigrams within a document collection. Here, bigrams refer to pairs of words that co-occur in the same context. The more times two words co-occur, the more semantically related they are, and the more likely they are to belong to the same topic. Based on this understanding, it is assumed that each bigram is generated by the same topic, and topics are generated from a mixed distribution of topics defined across the entire corpus. During the extraction process, low-frequency words, stop words, and other factors can be removed. Furthermore, the distance between two words can be constrained. For example, for short texts, the distance between two words can be arbitrary, while for long texts, the distance between two words is typically set between 30 and 60.

[0146] Figure 6C A more detailed flow chart of block 610 is shown for determining the probability distribution of each word in a title, according to some embodiments of the present disclosure.

[0147] like Figure 6C As shown, block 610 may further include block 612 , in which the computing device 20 may tokenize the title of the news text to obtain a plurality of words of the title.

[0148] At block 614, computing device 20 may organize two words from the plurality of words into a two-word pair and assign a topic from a predetermined topic set to the two-word pair. Here, the predetermined topic set may include, for example, a plurality of topic categories determined based on historical news text.

[0149] At block 616 , computing device 20 may determine a probability distribution of topics for all two-word pairs comprised of the plurality of words.

[0150] At block 618 , computing device 20 may determine a probability distribution for each word in each big-word pair based on the probability distribution for the topic of the big-word pair.

[0151] continue Figure 6A At block 620, at the topic clustering layer 652, computing device 20 may use a clustering algorithm to determine the topic category of the news text based on the probability distribution of the news text title within a predetermined topic set. In some embodiments, the clustering algorithm used may be the Xmeans algorithm, which does not require a specified number of clusters but instead iteratively selects the optimal number of clusters based on a clustering quality indicator.

[0152] Figure 6DA more detailed flow chart of block 620 for determining topic categories of news text using a clustering algorithm according to some embodiments of the present disclosure is shown.

[0153] like Figure 6D As shown, block 620 may further include block 622, where computing device 20 may cluster the news headlines into a plurality of clusters. For example, the news headlines may be clustered using an Xmeans algorithm.

[0154] In block 624, the computing device 20 may use the entity model to identify the mentioned entity in the news text. Here, the method of using the entity model to identify the mentioned entity in the news text may be, for example, as follows: Figures 7A to 7F Method 700 is shown, but the present disclosure is not limited in this regard.

[0155] Then, in block 626, the computing device 20 may perform similar cluster merging on the plurality of clusters based on the mentioned entities in the news text to determine the subject category of the news text. Here, similar cluster merging refers to merging clusters with a larger number of mentioned entity distributions into one cluster. In some implementations, the number of mentioned entities in each cluster may be counted to determine the core entities in each cluster (e.g., mentioned entities with a number greater than a certain threshold), and clusters with an overlap greater than a certain threshold may be merged into one cluster as a result of similar cluster merging of the mentioned entities in the news text.

[0156] continue Figure 6A At block 630, in the topic description layer 653, the computing device 20 may select central keywords from the news text to form a topic description for the news text. This way, the resulting topic description not only effectively reflects the topic of the news text but also provides a better textual presentation, making it easier to read and improving the user experience.

[0157] Figure 6E A more detailed flow chart of block 630 for determining a topic expression of a news text according to some embodiments of the present disclosure is shown.

[0158] like Figure 6E As shown, block 630 may further include block 632, where the computing device 20 may perform summary extraction on the news text. For example, in some embodiments, the computing device 20 may utilize the TextRank algorithm to perform summary extraction.

[0159] In block 634 , the computing device 20 may select a central keyword from the summary as a description of the topic based on the entity mentions in the news text and the word count of the news text.

[0160] also, Figures 6A to 6EThe method 600 shown may further include a historical topic merging process, which is used to merge historical topics based on similarity to prevent the predetermined topic set from being too large or dispersed.

[0161] More specifically, the historical topic merging process may include obtaining multiple historical topic descriptions in the predetermined topic set, that is, topic descriptions of multiple topics in the predetermined topic set. Then, a text similarity calculation may be performed on the multiple historical topic descriptions obtained to determine the text similarity between these historical topic descriptions. For example, TF-IDF (term frequency–inverse document frequency) may be used to calculate the cosine similarity between two historical topic descriptions as the text similarity. After determining the text similarity between these historical topic descriptions, the topics in the predetermined topic set may be merged based on these text similarities to update the predetermined topic set. For example, the text similarity may be compared with a predetermined similarity threshold, and two historical topics whose text similarity is less than the similarity threshold may be merged into one topic.

[0162] use Figures 6A to 6E The method shown here uses a topic classification model to determine the topic category of a news text. This allows for an assessment of the risk a news text poses to an enterprise based on whether the news text's topic category falls within a specific set of risk topics. Furthermore, the topic category of a news text can be used as an attribute of the public opinion entity it represents, facilitating further associations between public opinion entities and entities in the enterprise knowledge graph.

[0163] Figure 7A A flowchart of a method 700 for identifying entity mentions in news text using an entity model according to an embodiment of the present disclosure is shown.

[0164] like Figure 7A As shown in , method 700 may include block 710, where the computing device 20 may construct an enterprise knowledge graph for the enterprise. Figure 8A As shown above, the enterprise knowledge graph 810 may include multiple entities E1-E9 indicating the business elements of the enterprise, entity attributes of each entity and entity relationships between the entities (in Figure 8A (represented by lines between entities).

[0165] Here, the enterprise knowledge graph can be constructed based on the industry knowledge graph of the industry to which the enterprise belongs. Figure 8B A schematic diagram of an industry knowledge graph 820 according to an embodiment of the present disclosure is shown.

[0166] In this case, block 710 may further include: constructing an industry knowledge graph 820 based on the industry characteristics of the industry to which the enterprise belongs. The industry knowledge graph includes an entity set {E1}-{E9} indicating various business elements of the industry, an entity attribute set of each entity, and an entity relationship set between the entities.

[0167] For example, for the food industry, it focuses on business elements such as food categories, products, food materials, ingredients, enterprises, medicines, subordinate units, brands, nutrients, etc., so it can construct business elements such as Figure 8B For example, in the industry knowledge graph 820 shown, {E1} is an entity set representing brands, {E2} is an entity set representing products, {E3} is an entity set representing food categories, {E4} is an entity set representing ingredients, {E5} is an entity set representing nutrients, {E7} is an entity set representing enterprises, {E6} is an entity set representing enterprises' subsidiaries, {E8} is an entity set representing pharmaceuticals, and {E9} is an entity set representing food materials. Table 2 shows examples of various entities and their corresponding attribute sets.

[0168] Table 2

[0169]

[0170] The relationship between entities is shown in Table 3 below.

[0171] Table 3

[0172] relation entity Food material corresponding brand Food ingredients, brands The product contains food materials Products, food materials Food materials belong to the category Food materials, categories Food materials including ingredients Food materials, ingredients Food materials contain nutrients Food materials, nutrients Enterprises supply food materials Enterprise, food materials The enterprise has subordinate units Enterprises, subordinate units Enterprises supply drugs Enterprise, pharmaceutical

[0173] After constructing the industry knowledge graph 820, an enterprise knowledge graph 810 for the target enterprise can be constructed based on the target enterprise's business data and the industry knowledge graph 820. For example, a specific name in the entity set {E7} in the industry knowledge graph 820 can be selected as entity E7 in the enterprise knowledge graph 810, and a relationship between entity E7 and entity E8 can be established as "Enterprise supplies pharmaceuticals," and a relationship between entity E7 and entity E9 can be established as "Enterprise supplies food materials."

[0174] In addition, as mentioned above, the enterprise knowledge graph 810 may also include the public opinion entity corresponding to the news text and the mentioned entity identified from the news text. Specifically, a public opinion entity may be constructed for the news text, the attributes of which include the subject category and sentiment polarity of the news text, and the public opinion entity and the mentioned entity identified from the news text are added to the enterprise knowledge graph 810, such as Figure 8A The public opinion entity is shown in Table 4 below, where the attributes of the public opinion entity may include the topic category and sentiment polarity of the news text.

[0175] Table 4

[0176]

[0177] In addition, when there are multiple news texts, similarity relationships can also be established between the public opinion entities corresponding to each news text. For example, the vector distance between the public opinion entity and an existing public opinion entity can be determined, and when the vector distance is below a certain threshold, the two public opinion entities are considered similar, thereby establishing the relationship between the two public opinion entities as similar. For example, Table 5 shows two public opinion entities with similar relationships.

[0178] Table 5

[0179] relation entity resemblance Public opinion entity 1, public opinion entity 2

[0180] Here, as mentioned above, the method for determining the subject category of the news text may be, for example, combining Figures 6A to 6E The method 600 described above, the method for determining the sentiment polarity of the news text may be, for example, the method 500 described in conjunction with FIG5 . In addition, the method for identifying the mentioned entity from the news text may be, for example, the following method combined with Figures 7B to 7E and as described in detail in block 720.

[0181] In box 720, the computing device 20 can use the entity model to identify one or more mentioned entities, the attributes of each mentioned entity, and the entity relationships between the mentioned entities from the news text. Here, the description and attributes of the mentioned entities are shown in Table 2 above, and the entity relationships between the mentioned entities are shown in Table 3 above. The identification of entities and entity relationships can include a serial method or a joint identification method. For the former, the entity model can include two serial models, namely, an entity recognition model and an entity relationship model, which can first identify entities from the text and then identify entity relationships between entity pairs. The method of identifying entities and entity relationships in a serial manner is, for example, combined with the following Figure 7B As shown in detail. For the latter, entities and relationships between entities can be jointly extracted. For the joint extraction method, it can further include joint extraction based on parameter sharing and joint extraction of entity relationships based on joint decoding. The joint extraction method based on parameter sharing is for example combined with the following Figure 7C As shown in detail.

[0182] At block 730, the computing device 20 may determine the entity distance between the mentioned entity identified at block 720 and multiple entities in the enterprise knowledge graph (e.g., entities E1-E9 in the enterprise knowledge graph 810). The process of determining the entity distance may be, for example, as follows: Figure 7F As detailed.

[0183] At block 740, the computing device 20 may determine, based on the entity distance, that the mentioned entity hits the enterprise knowledge graph 810. Specifically, the computing device 20 may determine whether the entity distance calculated at block 730 is less than a predetermined distance threshold, and if the entity distance is less than the predetermined distance threshold, determine that the mentioned entity hits an entity in the enterprise knowledge graph 810. In this case, the entity in the enterprise knowledge graph 810 that is hit is also referred to as a hit entity.

[0184] Figure 7B FIG. 7 is a further detailed flowchart of block 720 for identifying entity mentions and entity relationships from news text using an entity model according to some embodiments of the present disclosure. Figure 7B In the illustrated embodiment, the entity model includes a serial entity recognition model and an entity relationship model, which work in a pipeline manner. First, the entity recognition model is used to identify mentioned entities from the news text, and then the entity relationship model is used to determine the entity relationship between the identified mentioned entities.

[0185] like Figure 7B As shown, box 720 may include box 721, in which the computing device 20 may annotate entities in the training text and entity relationships between entities. Here, the training text may be a historical news text in a historical news text library. Among them, entities that need attention, such as entities related to risk warnings for food safety, may be annotated, and entity relationships between these entities that need attention may be annotated, for example, based on at least a part of the entity relationships shown in Table 3, the entity relationships in the training text may be annotated. That is, both entity labels and entity relationship labels may be set in the training text. In addition, entity attributes may be selected based on the focus, for example, the attributes of interest may be selected from the entity attributes shown in Table 2, and the attributes of the entity may be annotated. Further, in some embodiments, the sentiment polarity at the entity level may also be annotated.

[0186] In block 722 , the computing device 20 may train the entity recognition model using the annotated training text to generate a trained entity recognition model.

[0187] In block 723, the computing device 20 may train the entity relationship model using the annotated training text to generate a trained entity relationship model. In this way, by annotating the entities and entity relationships in the training text, a trained entity recognition model and an entity relationship model are obtained using the training text.

[0188] In block 724 , the computing device 20 may utilize the trained entity recognition model to identify one or more mentioned entities and attributes of each mentioned entity from the news text to be recognized.

[0189] Depending on the entity labels used, the entity content that a trained entity recognition model can recognize will also vary. For example, entity labels typically include location and type, so the entity mentions that a trained entity recognition model can recognize will typically also include the location and type of the entity.

[0190] In block 725, the computing device 20 may utilize the trained entity relationship model and the one or more entity mentions identified from the news text in block 724 to determine the entity relationships between the mentioned entities. Depending on the entity relationship labels that are annotated, the entity relationships that the trained entity relationship model can identify may also vary. For example, if only the entity relationships between an enterprise and its subordinate units, as shown in Table 3, are annotated, the trained entity relationship model will only be able to identify relationships between the two when the mentioned entities are an enterprise and its subordinate units.

[0191] In addition, in some embodiments, for example, during a cold start, a keyword dictionary can be directly used to identify mentioned entities in the news text. For example, a string matching method can be used to directly search for keywords preset by the enterprise in the news text.

[0192] Figure 7C FIG. 7 shows a further detailed flow chart of block 720 for identifying entity mentions and entity relationships from news text using an entity model according to some other embodiments of the present disclosure. Figure 7C In the illustrated embodiment, the entity model includes a joint recognition model for entities and entity relationships. From the perspective of task completion, the joint recognition model still has to complete two tasks: entity recognition and entity relationship recognition, that is, it involves two sub-models. The difference is that in the joint recognition model, at least part of the two sub-models can be shared. For example, it can be entity relationship extraction based on parameter sharing or entity relationship extraction based on joint decoding.

[0193] like Figure 7C As shown in , box 720 may include box 721 ′, where the computing device 20 may label entities in the training text and entity relationships between entities.

[0194] Similarly, the training text can be a historical news text in a historical news text library. Entities that require attention, such as entities related to food safety risk warnings, can be annotated. Furthermore, entity attributes can be selected based on the focus, such as selecting the attributes of interest from the entity attributes shown in Table 2 and annotating the entity attributes. Furthermore, entity relationships in the training text can be annotated based on at least a portion of the entity relationships shown in Table 3. Furthermore, in some embodiments, entity-level sentiment polarity can also be annotated.

[0195] In block 722 ′, the computing device 20 may train the joint recognition model using the annotated training text to generate a trained joint recognition model.

[0196] In block 723 ′, the computing device 20 may use the trained joint recognition model to identify one or more mentioned entities, attributes of each mentioned entity, and entity relationships between the mentioned entities from the news text.

[0197] In some embodiments, the joint recognition model is an entity and relationship joint recognition model based on parameter sharing. Figure 7D A structural diagram of an entity and relationship joint recognition model 750 based on parameter sharing according to an embodiment of the present disclosure is shown. Figure 7E FIG. 7 shows a detailed flow chart of block 722′ for training the joint recognition model 750 using the annotated training text according to some embodiments of the present disclosure. Figure 7D As shown, the entity and relationship joint recognition model 750 based on parameter sharing may include an embedding layer 751, a shared encoding layer 752, an entity decoding layer 753, a relationship decoding layer 754 and a loss function layer 755.

[0198] like Figure 7E As shown in , box 722' further includes box 7221', where the computing device 20 can input the annotated training text into the embedding layer 751 to convert it into an embedding vector.

[0199] In block 7222', the computing device 20 may extract the semantic encoding of the training text from the embedding vector obtained in block 7221' using the context model in the shared encoding layer 752. Here, the semantic encoding may include grammatical and syntactic features of the training text.

[0200] In block 7223 ′, the computing device 20 may calculate the score distribution of the entity tag categories of the training text based on the semantic encoding obtained in block 7222 ′ at the entity decoding layer 753 .

[0201] In block 7224 ′, the computing device 20 may calculate the score distribution of the relationship tag categories of the training text based on the semantic encoding obtained in block 7222 ′ at the relationship decoding layer 754 .

[0202] In box 7225', the computing device 20 can use the cross entropy function in the loss function layer 755 to determine the loss value based on the score distribution of the entity label category, the score distribution of the relationship label category and the true distribution of the training text.

[0203] At block 7226′, the computing device 20 may iteratively optimize the joint recognition model 750 based on the loss value determined at block 7225′ to obtain a trained joint recognition model. Here, a backpropagation algorithm may be used to update the model parameters of the joint recognition model 750 to implement the dependency between the entity decoding layer 753 and the relationship decoding layer 754.

[0204] In some embodiments, the shared encoding layer 752 may include a bidirectional LSTM model, the entity decoding layer 753 may include an LSTM model, and the relationship decoding layer 754 may include a CNN model. In other words, the joint recognition model 750 is a joint recognition model based on LSTM and CNN.

[0205] Utilizing the joint recognition model 750, the label dependency problem in entity recognition is solved by encoding the input text through a shared encoding layer and then using an LSTM model and a CNN model for entity recognition and entity relationship classification, respectively. In addition, a backpropagation algorithm is used to update the shared parameters during iterative optimization of the model to realize the dependency between the two subtasks.

[0206] Figure 7F A more detailed flow chart of block 730 for determining entity distance is shown in accordance with an embodiment of the present disclosure.

[0207] like Figure 7F As shown, block 730 may include block 732, where the computing device 20 may determine a word vector as an entity name distance based on the entity name of the mentioned entity determined in block 720 and the entity name of an entity in the enterprise knowledge graph. Figure 8A As shown in , the mentioned entity A1 and the entity E2 in the enterprise knowledge graph 810 can be represented as word vectors V A1 and V E2 (The word vector can be the word vector after removing non-keyword characters.) Then, the entity name distance between the two entities can be expressed as D1 = || V A1 -V E2 ||.

[0208] At block 734, the computing device 20 may determine multiple word vectors as entity attribute distances based on each attribute of the mentioned entity and multiple attributes of an entity in the enterprise knowledge graph. Figure 8A As shown in , the attributes of the mentioned entity A1 and the attributes of the entity E2 in the enterprise knowledge graph 810 can be represented as multiple word vectors V A1_i and V E2_i Then, the entity attribute distance between these two entities can be expressed as

[0209]

[0210] At block 736, the computing device 20 may determine the entity distance between the mentioned entity A1 and the entity E2 in the enterprise knowledge graph 810 based on the entity name distance D1 and the entity attribute distance D2. In some embodiments, the entity distance D = D1 + D2.

[0211] In some other embodiments, the entity distance D=α1*D1+α2*D2, where α1 and α2 are empirical weight values.

[0212] Use Figures 7A to 7F The method 700 shown, by using an entity model to identify entities in news texts and determining whether the identified entities hit entities in the enterprise knowledge graph, can determine from the entity level whether the news text poses a public opinion risk to the enterprise.

[0213] Figure 9 FIG2 is a block diagram of a computing device 900 suitable for implementing embodiments of the present disclosure. The computing device 900 may be, for example, the computing device 20 described above.

[0214] like Figure 9 As shown in the figure, the computing device 900 may include one or more central processing units (CPUs) 910 (only one is schematically shown in the figure), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) 920 or computer program instructions loaded from a storage unit 980 into a random access memory (RAM) 930. Various programs and data required for the operation of the computing device 900 may also be stored in the RAM 930. The CPU 910, ROM 920, and RAM 930 are connected to each other via a bus 940. An input / output (I / O) interface 950 is also connected to the bus 940.

[0215] Various components in the computing device 900 are connected to the I / O interface 950, including an input unit 960, such as a keyboard, a mouse, etc.; an output unit 970, such as various types of displays, speakers, etc.; a storage unit 980, such as a magnetic disk, an optical disk, etc.; and a communication unit 990, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 990 allows the computing device 900 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0216] The methods 200 to 700 described above may be executed, for example, by the CPU 910 of the computing device 900. For example, in some embodiments, the methods 200 to 700 may be implemented as a computer software program that is tangibly included in a machine-readable medium, such as the storage unit 980. In some embodiments, part or all of the computer program may be loaded and / or installed on the computing device 900 via the ROM 920 and / or the communication unit 990. When the computer program is loaded into the RAM 930 and executed by the CPU 910, one or more operations of the methods 200 to 700 described above may be performed. In addition, the communication unit 990 may support wired or wireless communication capabilities.

[0217] Those skilled in the art will understand that Figure 9 The computing device 900 shown is for illustration only. In some embodiments, the computing device 20 may include more or fewer components than the computing device 900.

[0218] The above description of the methods 200 and 300 for performing enterprise risk assessment based on news text according to the present disclosure, as well as specific implementation methods 400 to 700 of a portion of these methods, and a computing device 900 that can be used as the computing device 20 are described in conjunction with the accompanying drawings. However, those skilled in the art will appreciate that the execution of the steps and sub-steps of the methods 200 to 700 is not limited to the order shown in the drawings and described above, but can be executed in any other reasonable order. In addition, the computing device 900 does not necessarily include Figure 9 All components shown in the figure may include only some of the components necessary to perform the functions described in the present disclosure, and the connection method of these components is not limited to the form shown in the figure.

[0219] The present disclosure may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present disclosure.

[0220] In one or more exemplary designs, the functions described in this disclosure may be implemented using hardware, software, firmware, or any combination thereof. For example, if implemented using software, the functions may be stored as one or more instructions or codes on a computer-readable medium, or transmitted as one or more instructions or codes on a computer-readable medium.

[0221] The various units of the apparatus disclosed herein may be implemented using discrete hardware components or may be integrated into a single hardware component, such as a processor. For example, the various exemplary logic blocks, modules, and circuits described in conjunction with the present disclosure may be implemented or executed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof for performing the functions described herein.

[0222] Those skilled in the art should also understand that the various illustrative logical blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments of the present disclosure may be implemented as electronic hardware, computer software, or a combination of both.

[0223] The above description of the present disclosure is intended to enable any person skilled in the art to implement or use the present disclosure. Various modifications of the present disclosure will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit and scope of the present disclosure. Therefore, the present disclosure is not limited to the examples and designs described herein, but is intended to be consistent with the widest scope of the principles and novel features disclosed herein.

Claims

1. A method for enterprise risk assessment based on news text, comprising: Determining whether the news text meets a predetermined condition; as well as In response to determining that the news text satisfies the predetermined condition, determining that the enterprise has a public opinion risk, wherein determining that the news text satisfies the predetermined condition includes at least two of the following: Determining that the event risk value of the public opinion event corresponding to the news text is greater than a risk threshold; Determining that the sentiment polarity of the news text is lower than a sentiment threshold; Determining a topic category of the news text and determining that the determined topic category belongs to a specific risk topic set; as well as Identify the mentioned entity in the news text and determine whether the mentioned entity hits the enterprise knowledge graph of the enterprise, Determining the event risk value of the public opinion event corresponding to the news text includes: Establish a public opinion event model to at least identify event types, event trigger words corresponding to each event type, and event elements; Identifying a public opinion event from the news text using the public opinion event model, where the public opinion event includes at least a trigger word and an event element of the public opinion event; and Use the event risk model to determine the event risk value of the public opinion event. The event risk model is used to determine the event risk value of the public opinion event, including: Constructing a feature word set of the public opinion event based on the event type, event trigger words and event elements of the public opinion event; Determining an average word vector of the feature word set; comparing the average word vector with a plurality of average word vectors in the parameters of the event risk model to determine a closest average word vector; and The risk value of the public opinion event is determined based on the closest average word vector.

2. The method according to claim 1, wherein identifying public opinion events from the news text using the public opinion event model comprises: Determining an event representation mode of the news text, wherein the event representation mode includes any one of a headline event representation, a central event representation, an event extraction representation, and an event semantic representation; Based on the event representation method, the public opinion event is identified from the news text using the public opinion event model; as well as Generate a short text public opinion event representation for the identified public opinion event.

3. The method as claimed in claim 1, wherein the public opinion event model is further used to identify event element roles indicating the roles of the event elements and / or event description sentences containing event trigger words and event elements.

4. The method of claim 1 , further comprising: Building a historical event library associated with the enterprise, the historical event library containing multiple historical events of the risk type of interest, each historical event being built based on the public opinion event model; Marking each historical event in the historical event database with a risk value, wherein the risk value indicates the risk level of the historical event; Training the event risk model using historical events with risk values ​​to generate a trained event risk model; as well as The public opinion event is input into a trained event risk model to determine an event risk value of the public opinion event.

5. The method of claim 1 , wherein determining that the sentiment polarity of the news text is lower than a sentiment threshold comprises: Determining a sentiment polarity of the news text using a sentiment classification model, the sentiment polarity indicating positive sentiment, negative sentiment, or neutral sentiment; as well as The sentiment polarity of the news text is compared with the sentiment threshold to determine whether the sentiment polarity of the news text is lower than the sentiment threshold.

6. The method of claim 5, wherein determining the sentiment polarity of the news text using a sentiment classification model comprises: Filtering text data contained in the news text; Segmenting the filtered text data, and mapping the segmented text data into text identifier sequences, wherein each text identifier sequence includes one or more discrete encoding vectors; Inputting the generated text identifier sequence into the embedding layer of the sentiment classification model to convert the text identifier sequence into a continuous embedding vector; Encoding the embedding vector output by the embedding layer at the encoding layer of the sentiment classification model to generate a contextual representation of the embedding vector; Generating a sentence-level feature vector of the embedding vector using an attention function in the pooling layer of the sentiment classification model; generating a quality probability distribution of the sentence-level feature vector, wherein the quality probability distribution indicates the probability of the news text being at different sentiment polarities; and The sentiment polarity with the highest probability in the quality probability distribution is selected as the sentiment polarity of the news text.

7. The method of claim 5, further comprising: Constructing an emotion training text set associated with the enterprise, wherein the emotion training text set includes a plurality of emotion training texts, and each emotion training text includes a news text associated with the enterprise; Labeling the sentiment polarity of each sentiment training text in the sentiment training text set; as well as The sentiment classification model is trained using the sentiment training text and the marked sentiment polarity to generate a trained sentiment classification model.

8. The method according to claim 7, wherein labeling the sentiment polarity of each sentiment training text in the sentiment training text set comprises: Searching for risk words in a risk word dictionary in the sentiment training text; If a risk word is found in the sentiment training text, determining the semantics associated with the risk word based on dependency syntactic rules; as well as The sentiment training text is annotated based on the semantics associated with the risk words.

9. The method of claim 1 , wherein determining the topic category of the news text comprises determining the topic category of the news text using a topic classification model, the topic classification model comprising a topic extraction layer and a topic clustering layer, and determining the topic category of the news text using the topic classification model comprises: At the topic extraction layer, using a two-word topic model, determining the probability distribution of the title of the news text in a predetermined topic set and the probability distribution of each word in the title in each topic in the predetermined topic set; and At the topic clustering layer, a clustering algorithm is used to determine the topic category of the news text based on the probability distribution of the title of the news text in a predetermined topic set.

10. The method according to claim 9, wherein the topic classification model further comprises a topic description layer, and the method further comprises: At the topic description layer, central keywords are selected from the news text to form a topic description of the news text.

11. The method of claim 10, wherein at the topic extraction layer, using a two-word topic model, determining the probability distribution of the title of the news text in a predetermined topic set and the probability distribution of each word in the title in each topic in the predetermined topic set comprises: Segmenting the title of the news text to obtain a plurality of words of the title; organizing two words from the plurality of words into a bigram pair and assigning the bigram pair to a topic from the predetermined topic set; Determining a probability distribution of topics of all two-word pairs formed by the plurality of words; as well as A probability distribution for each word in each big-word pair is determined based on the probability distribution of the subject of the big-word pair.

12. The method of claim 11, wherein at the topic clustering layer, determining the topic category of the news text using a clustering algorithm based on the probability distribution of the title of the news text in a predetermined topic set comprises: Cluster news headlines into multiple clusters; identifying mentioned entities in the news text using an entity model; as well as The plurality of clusters are similarly clustered and merged based on the mentioned entities in the news text to determine a topic category of the news text.

13. The method according to claim 11, wherein at the topic description layer, selecting central keywords from the news text to form a topic description of the news text comprises: Extracting a summary of the news text; as well as Based on the mentioned entities in the news text and the word count of the news text, a central keyword is selected from the summary as the topic description.

14. The method of claim 11, further comprising: Obtaining a plurality of historical topic descriptions in the predetermined topic set; Performing text similarity calculation on the plurality of historical topic descriptions to determine text similarities between the plurality of historical topic descriptions; and Based on the text similarities between the multiple historical topic descriptions, topics in the predetermined topic set are merged to update the predetermined topic set.

15. The method of claim 1, wherein identifying the mentioned entity in the news text using an entity model and determining that the mentioned entity hits the enterprise knowledge graph of the enterprise comprises: Constructing an enterprise knowledge graph of the enterprise, the enterprise knowledge graph including a plurality of entities indicating business elements of the enterprise, entity attributes of each entity, and entity relationships between the entities; Identifying one or more mentioned entities, attributes of each mentioned entity, and entity relationships between the mentioned entities from the news text using the entity model; Determining entity distances between the mentioned entity and multiple entities in the enterprise knowledge graph; as well as Determine whether the mentioned entity hits the enterprise knowledge graph based on the entity distance.

16. The method according to claim 15, wherein constructing the enterprise knowledge graph of the enterprise comprises: Building an industry knowledge graph based on the industry characteristics of the industry to which the enterprise belongs, wherein the industry knowledge graph includes a set of entities indicating various business elements of the industry, a set of entity attributes of each entity, and a set of entity relationships between the entities; and Construct an enterprise knowledge graph of the enterprise based on the business data of the enterprise and the industry knowledge graph.

17. The method of claim 16, further comprising: Constructing a public opinion entity for the news text, wherein the attributes of the public opinion entity include the subject category and sentiment polarity of the news text; Add the public opinion entity and the mentioned entity to the enterprise knowledge graph.

18. The method of claim 16, wherein the entity model comprises a serial entity recognition model and an entity relationship model, and the method further comprises: Label the entities in the training text and the entity relationships between them; Training the entity recognition model using the annotated training text to generate a trained entity recognition model; Training the entity relationship model using the annotated training text to generate a trained entity relationship model; Using the trained entity recognition model, identifying one or more mentioned entities and attributes of each mentioned entity from the news text; Utilizing the trained entity relationship model and one or more mentioned entities identified from the news text, entity relationships between the mentioned entities are determined.

19. The method of claim 16, wherein the entity model comprises a joint recognition model of entities and entity relationships, and the method further comprises: Label the entities in the training text and the entity relationships between entities; Training the joint recognition model using the annotated training text to generate a trained joint recognition model; as well as The trained joint recognition model is used to identify one or more mentioned entities, attributes of each mentioned entity, and entity relationships between the mentioned entities from the news text.

20. The method of claim 19, wherein the joint recognition model is an entity and relationship joint recognition model based on parameter sharing.

21. The method of claim 20, wherein the parameter-sharing-based entity and relationship joint recognition model comprises an embedding layer, a shared encoding layer, an entity decoding layer, a relationship decoding layer, and a loss function layer, and training the joint recognition model using the annotated training text to generate a trained joint recognition model comprises: Inputting the annotated training text into the embedding layer to convert it into an embedding vector; Extracting a semantic code of the training text from the embedding vector using a context model at the shared coding layer, wherein the semantic code includes grammatical and syntactic features of the training text; At the entity decoding layer, calculating the score distribution of the entity label categories of the training text according to the semantic encoding; At the relation decoding layer, calculating the score distribution of the relation label categories of the training text according to the semantic encoding; In the loss function layer, a cross entropy function is used to determine a loss value based on the score distribution of the entity label category, the score distribution of the relationship label category, and the true distribution of the training text; as well as The parameter-sharing-based entity and relationship joint recognition model is iteratively optimized based on the loss value.

22. The method of claim 15, wherein the attribute of the entity mention comprises a sentiment polarity of the entity mention.

23. The method of claim 15, wherein determining the entity distance between the mentioned entity and a plurality of entities in the enterprise knowledge graph comprises: Determine a word vector as an entity name distance based on the entity name of the mentioned entity and the entity name of an entity in the enterprise knowledge graph; Determining multiple word vectors as entity attribute distances based on each attribute of the mentioned entity and multiple attributes of an entity in the enterprise knowledge graph; and An entity distance between the mentioned entity and an entity in the enterprise knowledge graph is determined based on the entity name distance and the entity attribute distance.

24. The method of claim 23, wherein determining that the mentioned entity hits the enterprise knowledge graph based on the entity distance comprises: determining whether the entity distance is less than a predetermined distance threshold; as well as If the entity distance is less than the predetermined distance threshold, it is determined that the mentioned entity hits the entity in the enterprise knowledge graph.

25. A computing device comprising: at least one processor; as well as At least one memory coupled to the at least one processor and storing instructions for execution by the at least one processor, the instructions, when executed by the at least one processor, causing the computing device to perform the steps of the method according to any one of claims 1 to 24.

26. A computer-readable storage medium having computer program code stored thereon, wherein the computer program code, when executed, performs the method according to any one of claims 1 to 24.

Images